30-meter Land Surface Temperature from Landsat via Progressive Self-Training Downscaling

Abstract: Land surface temperature (LST) is a critical parameter for characterizing surface energy balance and hydrothermal processes. While Landsat provides invaluable LST observations at medium spatial resolution for over 40 years, its native spatial resolution of thermal bands (e.g., 100 m) remains insufficient compared to its 30 m optical bands, failing to meet the demands of fine-scale studies. To address this issues, this study proposes a progressive self-training framework for downscaling Landsat LST to 30 m without relying on fine-scale ground truth, while maintaining minimal data dependence. The framework progressively optimizes a cross-modal fusion network to refine thermal details in a coarse-to-fine manner, characterized by one pre-training and two fine-tuning stages. Spatial validation against SDGSAT-1 30 m LST and temporal validation using in situ measurements confirm its reliability and accuracy, with both station-averaged MAE and RMSE outperforming the official cubic product by approximately 0.4 K. Further performance comparison experiments demonstrate that the proposed framework consistently reconstructs coherent fine-scale thermal patterns while preserving spatial heterogeneity. Multi spatial resolution evaluations and ablation studies verify the effectiveness of the proposed strategy and network design. Overall, the framework provides a stable pathway for enhancing the spatial resolution of Landsat LST, providing fine-resolution data support for fine-scale surface process studies and localized environmental monitoring.

• The paper presents a three-stage self-training framework that downscales Landsat TIR data to 30m LST without relying on fine-scale ground truth.
• It integrates novel modules like cross-modal dictionary attention fusion, a hybrid CNN-Transformer backbone, and frequency-domain guidance to improve spatial reconstruction and accuracy.
• Validation demonstrates reductions in MAE, RMSE, and improved structural congruence over conventional methods, supporting enhanced environmental monitoring and urban heat analyses.

Progressive Self-Training Downscaling of Landsat Land Surface Temperature to 30 Meters

Introduction and Motivation

Land surface temperature (LST) is fundamental for characterizing terrestrial energy and hydrological processes across spatiotemporal scales. Although the Landsat satellite archive offers four decades of high-quality thermal infrared (TIR) data, the coarse resolution of the native thermal bands (100–120 m for Landsat 8/9) constrains their utility in applications requiring fine-grained thermal mapping, such as urban microclimate assessment, precision agriculture, and heterogeneous landscape analyses. Conventional LST downscaling methodologies—spatiotemporal fusion, multi-parameter fusion, and machine/deep learning—are impeded by strict requirements for ground-truth at fine spatial scales or by strong assumptions of scale-invariance between LST and auxiliary variables. Most notably, fine-resolution ground-truth LST is typically unavailable, and statistical relationships with auxiliary predictors (e.g., NDVI, NDBI, DEM) are distinctly scale-dependent, leading to model uncertainty and limited applicability.

Methodology

A dedicated progressive self-training downscaling framework is introduced to address the lack of fine-resolution LST ground truth and the intrinsic scale-dependence in LST-auxiliary variable relationships. The central advance is a multi-stage, coarse-to-fine transfer process, in which the model is (i) pre-trained at a coarse scale using available satellite products, and (ii) fine-tuned in subsequent stages guided by pseudo-labels generated from earlier stages and finer-resolution auxiliary data. This approach eschews any dependence on external high-resolution ground-truth measurements, utilizing only public Landsat products and auxiliary data (NDVI, NDWI, NDBI, DEM).

The cross-modal fusion network (CFDN-LST) integrates several architectural innovations:

• Cross-Modal Dictionary Attention Fusion (CDAF): This module learns a dictionary-based shared feature space for enhanced fusion of coarse-resolution LST and fine-resolution auxiliary data, disentangling shared and modality-unique structure.
• Hybrid CNN-Transformer Backbone: Local spatial details are reconstructed through lightweight CNN modules (LCM), while the Lightweight Transformer Module (LTM) leverages global non-local dependencies for structural coherence.
• Multi-Scale Feature Extraction (MFE): Rich contextual patterns at multiple spatial scales are captured from auxiliary variables for improved representation.
• Frequency-Domain Guidance: Both high-frequency preservation blocks and a Laplacian-based high-frequency structural loss are incorporated, using NDVI as a structural reference. This dual guidance prevents artifact accumulation and enforces texture fidelity during progressive refinement.

Model training proceeds in three stages: (1) coarse resolution pre-training (240m → 120m), (2) first fine-tuning (120m → 60m), and (3) second fine-tuning (60m → 30m), utilizing staged pseudo-labels and auxiliary variables resampled accordingly.

Evaluation and Results

Validation is conducted using a combination of independent in situ measurements, SDGSAT-1 30m LST imagery, and the Chinese Land Cover Dataset (CLCD). The progressive self-training framework consistently outperformed the official cubic-convolution resampled Landsat LST and leading alternatives (TsHARP, LSTDRN), both in structural detail and radiometric fidelity.

Key quantitative outcomes include:

• Station-averaged MAE and RMSE reductions of ~0.4 K compared to cubic-convolved products.
• RMSE against SDGSAT-1 LST reference: 2.36 K (proposed) vs. 2.85 K (official cubic).
• Cross-entropy (structural congruence): 3.26 (proposed) vs. 4.04 (official).
• Average Bias/MAE/RMSE/R^2 against in situ: 0.11 K / 3.35 K / 4.05 K / 0.86 (proposed) versus 0.81 K / 3.71 K / 4.41 K / 0.83 (official).

The framework preserved fine-scale thermal heterogeneity and reconstructed sharper land cover boundaries, outperforming linear and standard deep learning methods in both simulated and real-world benchmarks across diverse surface types. Ablation studies verified the critical role of each auxiliary variable and network module, with especially strong degradation when removing the CDAF fusion or DEM components.

Discussion

The staged self-training paradigm effectively mitigates the absence of fine-scale ground-truth by leveraging the model’s own predictions as weak supervision and integrating high-resolution auxiliary data. The cross-modal fusion with dictionary attention addresses scale-dependent spectral-thermal associations, enhancing both interpretability and performance. However, the framework cannot fully eliminate the propagation of errors inherent in pseudo-labels, particularly in very heterogeneous landscapes, and the progressive multi-stage paradigm incurs additional complexity during both training and inference.

Theoretically, the framework’s reliance solely on Landsat and freely available DEM/reflectance data endows it with strong generalizability and operational practicality. Practically, it provides a robust tool for fine-scale environmental monitoring, energy balance modeling, and urban heat mapping where direct thermal measurements are unavailable. The methodology is extendable to even finer resolutions should higher-resolution auxiliary predictors become available, and could benefit from integration with high-temporal-frequency LST measurements to further improve temporal density.

Conclusion

This work presents a progressive self-training downscaling framework for generating 30m LST from Landsat TIR data without fine-resolution ground-truth supervision. Through staged pseudo-labeling, cross-modal dictionary-based feature fusion, and explicit frequency-domain guidance, the approach achieves superior reconstruction of fine-scale LST spatial structure and consistency with independent high-resolution observations. The framework establishes a practical pathway for enhancing the applicability of Landsat LST in fine-scale geophysical and environmental analysis and offers a foundation for future unified or temporally dense satellite LST downscaling frameworks.

Future work should focus on error accumulation control, reduction of training complexity, and joint spatiotemporal LST refinement using data streams with high temporal revisit.